search

themrzmaster / git-re-basin-pytorch

Git Re-Basin: Merging Models modulo Permutation Symmetries in PyTorch
MIT License
72 stars 9 forks source link

ResNet50 #5

Closed superkirill closed 2 years ago

superkirill commented 2 years ago

Could you also provide a permutation spec for ResNet50?

themrzmaster commented 2 years ago

done it. let me know if it works as expected, my GPU takes a while to train a resnet50. use depth 52 as an arg

superkirill commented 2 years ago

Thanks! My concern is that it's not a very "canonical" implemetation with 4 block groups and varying numbers of channels, so it's hard to use it with official PyTorch models

themrzmaster commented 2 years ago

Yes, it should be easy to use the same logic with resnet from torchvision. Just have to remember to not use batchnorm

santosh9sanjeev commented 4 months ago

Its not working with torchvisions ResNet50. Can you pls help me out Thanks!

adampan0527 commented 3 months ago

cifar100_resnet52_4_weight_matching_interp_accuracy_epoch Hi, whem I try ResNet50 on cifar100, the accuracy drops a lot, and permuted accuracy is even worse than naive which is unexpected.